A curved surface polishing method, device, equipment and computer readable storage medium

Abstract

The application discloses a curved surface polishing method, device, equipment and computer readable storage medium, comprising: acquiring point cloud data, visual image and user instruction, constructing a grid model based on the point cloud data, calculating the normal vector and curvature data of each vertex, and performing regional division, processing the normal vector, curvature data and region type of each vertex to obtain a curved surface semantic vector, inputting the visual image, user instruction and curved surface semantic vector into a model to obtain a polishing strategy, and controlling a robot to polish, collecting force data and vibration signals, determining an abnormal condition based on the force data, vibration signals, difference process parameters and abnormal threshold, adjusting the polishing strategy based on the abnormal condition, and returning to execute the polishing strategy to control the robot to polish until the polishing is completed. Based on the curved surface semantic features, the application generates a polishing strategy, and adjusts the polishing strategy in real time based on the state data in the polishing process, so that the robot realizes curved surface polishing.

Description

Technical Field

[0001] This application relates to the field of robotic polishing technology, and in particular to a method, apparatus, equipment and computer-readable storage medium for polishing curved surfaces. Background Technology

[0002] With the rapid development of high-end manufacturing industries such as aerospace, automobile manufacturing, high-speed rail, shipbuilding and mold processing, the surface processing quality of complex curved parts (such as aircraft fuselages, engine blades, automobile body panels, precision molds, etc.) directly affects the aerodynamic performance, fatigue life and appearance quality of products. Therefore, surface grinding has become an indispensable key process in the high-end manufacturing field.

[0003] Traditional surface grinding has long relied primarily on manual labor. Workers use handheld grinding tools, relying on experience to grind the workpiece surface point by point and area by area. While manual grinding offers some flexibility, it suffers from inconsistent grinding quality. Influenced by worker skill level, fatigue, and subjective judgment, it is prone to defects such as over-grinding, under-grinding, vibration marks, and burns, resulting in low product qualification rates, high rework rates, and low production efficiency. This makes it difficult to meet the demands of large-scale, high-precision manufacturing. For large, complex curved workpieces, manual grinding cannot guarantee uniformity of grinding posture and contact force, making it difficult to control processing consistency.

[0004] In recent years, robots have been widely used in industrial production due to their advantages such as high repeatability, ability to work continuously for long periods of time, and adaptability to harsh environments. They have provided important support for the automation of surface treatment processes. Therefore, how to control robots to polish curved surfaces has been a hot topic of concern. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, device and computer-readable storage medium for polishing curved surfaces, so as to control a robot to polish curved surfaces.

[0006] To achieve the above objectives, the following solution is proposed: A method for polishing curved surfaces, comprising: Acquire 3D point cloud data, visual image data, and user language commands for the workpiece to be polished; Based on the three-dimensional point cloud data, a triangular mesh digital surface model of the workpiece to be polished is constructed. Calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model; Based on the normal vectors and curvature data of each vertex, the triangular mesh digital surface model is divided into regions, and the region type label of each region is determined. By using a pre-trained lightweight graph neural network, the normal vectors, curvature data and region type labels of each vertex are encoded and mapped to a high-dimensional semantic embedding space to obtain structured surface semantic vectors. Based on the visual image data, the user language instructions, and the structured surface semantic vector, a pre-trained polishing strategy generation model is input to obtain a structured polishing strategy file, which contains differential process parameters corresponding to each region. Based on the structured grinding strategy file, the robot is controlled to grind the workpiece to be ground and an execution record is generated. Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process; Based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex, and the structured grinding strategy file, the robot end effector position is corrected in real time. Based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, abnormal situations are identified, and based on the abnormal situations, it is determined whether it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy. If necessary, based on the aforementioned abnormal situation and the structured surface semantic vector corresponding to the current region, similar adjustment decision records are searched from a pre-created historical experience case library; Based on the six-dimensional force signal data, the tactile vibration signal, the structured surface semantic vector corresponding to the current region, and the similar adjustment decision records, the prompt words are determined; The prompt word is input into the adjustment decision model to obtain the adjustment instruction. The adjustment decision model is configured to have the ability to obtain the current state context matrix based on the six-dimensional force signal data and the tactile vibration signal, obtain the anomaly cause reasoning result based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain the adjustment instruction based on the anomaly cause reasoning result and the similar adjustment decision record. Based on the adjustment instructions, the structured grinding strategy file is adjusted accordingly to obtain the adjusted structured grinding strategy instructions. Then, the process of controlling the robot to grind the workpiece based on the structured grinding strategy file is returned to be executed until the grinding is completed.

[0007] Optionally, constructing a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data includes: For each point, calculate the average distance from all points in its neighborhood to that point; Points whose average distance exceeds the preset global average distance threshold are identified as outliers and removed. The 3D point cloud space is divided into voxel grids of a specified size. The 3D point cloud data, after removing outlier noise, is simplified by retaining only one representative point in each grid. Based on the simplified 3D point cloud data, a continuous triangular mesh digital surface model of the workpiece to be polished is constructed.

[0008] Optionally, the curvature data includes average curvature, and the triangular mesh digital surface model is divided into regions based on the normal vectors and curvature data of each vertex, and the region type label of each region is determined, including: From the vertices whose region type labels have not been determined, select the vertex corresponding to the minimum average curvature as the current seed point, and determine the region type label of the region where the current seed point is located based on the normal vector and curvature data of the current seed point. The neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold are included in the region where the current seed point is located. Determine whether there are any newly added neighboring vertices in the region where the current seed point is located; If there are newly added neighboring vertices in the region where the current seed point is located, then the newly added neighboring vertices in the region where the current seed point is located are sequentially used as new current seed points, and the process returns to the step of adding the neighboring vertices with undetermined region type labels that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold to the region where the current seed point is located. If no new neighboring vertices are added to the region where the current seed point is located, then return to the step of selecting the vertex with the minimum average curvature from the vertices whose region type labels have not been determined as the current seed point.

[0009] Optionally, the polishing strategy generation model includes: an input layer, a multimodal semantic fusion layer, a polishing state understanding and process reasoning layer, a partitioning strategy and process parameter generation layer, a baseline polishing path generation layer, and a structured polishing strategy file output layer. The visual image data, the user language instructions, and the structured surface semantic vector are obtained through the input layer. The multimodal semantic fusion layer maps the input visual image data, user language commands, and structured surface semantic vectors to a unified semantic embedding space, thereby achieving adaptive fusion of cross-modal features and obtaining fused semantic features. Through the grinding state understanding and process reasoning layer, based on the fused semantic features, the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each area, and the current grinding stage are determined. Through the partitioning strategy and process parameter generation layer, the processing order between regions is generated based on the comprehensive processing difficulty level of each region and the spatial adjacency relationship between each region. Based on the processing order between regions and the fused semantic features, the differentiated process parameters corresponding to each region are generated. Through the aforementioned reference grinding path generation layer, for each region, based on the region type label corresponding to each region, different path generation strategies are adopted to create target grinding points corresponding to each region. According to the pre-set objective function, the smoothness of the target grinding points is optimized, and the grinding path corresponding to each region is determined. Based on the processing sequence between the regions and the grinding path corresponding to each region, the transition relationship between regions is determined. The target grinding point includes spatial coordinates, tool posture, desired contact force, and desired speed. The structured grinding strategy file output layer generates a structured grinding strategy file based on the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each region, the current grinding stage, the differentiated process parameters corresponding to each region, the path generation strategy, the target grinding point number corresponding to each region, and the transition relationship between regions.

[0010] Optionally, the differential process parameters include the target normal contact force, and the real-time correction of the robot's end effector posture position based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex, and the structured grinding strategy file includes: The normal vector of the current contact point is determined based on the normal vectors of the three vertices on the triangular surface where the current contact point is located. Determine the current normal contact force based on the six-dimensional force signal data and normal vector at the current contact point; The deviation between the current normal contact force and the target normal contact force is calculated in real time. Using the deviation value and the differential process parameters corresponding to the current region, calculate the robot end position correction amount, and based on the robot end position correction amount, correct the robot end position in the normal direction of the current contact point; The path tangent direction of the current contact point is determined based on the structured polishing strategy file; Based on the normal vector of the current contact point and the tangent direction of the path, the desired posture of the robot end effector is determined; Acquire the current posture data of the robot's end effector and calculate the current posture error; Based on the current attitude error and angular velocity error, a robot end-effector attitude correction command is generated, and the robot end-effector attitude is corrected in the tangential plane direction of the surface where the current contact point is located based on the robot end-effector attitude correction command.

[0011] Optionally, determining abnormal situations based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and a pre-set abnormal threshold includes: Based on the six-dimensional force signal data, a real-time force perception semantic vector is generated; Based on the tactile vibration signal, a vibration feature vector is extracted; By using force perception semantic vectors with preset periods and corresponding vibration feature vectors, and concatenating them according to time series, a state context matrix is ​​constructed. Based on the state context matrix, the differential process parameters corresponding to each region, and the pre-set anomaly threshold, multi-dimensional anomaly detection is performed to determine the abnormal situation.

[0012] Optionally, the state context matrix includes the actual normal contact force, actual vibration amplitude, actual root mean square value of acoustic emission, actual force signal variance, and actual grinding area. The differential process parameters include the target normal contact force. Based on the state context matrix, the differential process parameters corresponding to each region, and a pre-set anomaly threshold, multi-dimensional anomaly detection is performed to determine abnormal situations, including: If the difference between the actual normal contact force and the target normal contact force is greater than the preset abnormal threshold for normal contact force, and the duration is greater than the preset deviation time threshold, then it is determined to be a force deviation warning. If the rate of change of the difference between the actual normal contact force and the target normal contact force is greater than the preset threshold for the rate of change of the normal contact force deviation, it is determined as a force change warning. If the actual vibration amplitude is greater than the preset vibration amplitude threshold, it is determined to be a tremor risk warning; If the actual root mean square value of acoustic emission is greater than the preset root mean square value threshold of acoustic emission, it is determined as a tool passivation or burn warning. If the actual force signal variance is greater than the preset force signal variance, it is determined to be a contact instability warning; If the actual polishing area differs from the target polishing area, it will be considered a parameter update warning. Record the warning situation and the number of warnings as abnormal situations.

[0013] Optionally, after receiving the adjusted structured grinding strategy instructions and controlling the robot to grind the workpiece based on the adjusted structured grinding strategy file, the process further includes: Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process; Based on the six-dimensional force signal data and the tactile vibration signal, monitor the changing trend of the abnormal situation; If the abnormal indicators corresponding to the abnormal situation gradually converge to the target value, the adjustment is deemed effective and recorded as a successful experience. If the abnormal indicator corresponding to the abnormal situation does not gradually converge to the target value or a new abnormal situation occurs, the adjustment is determined to be invalid, the current adjustment instruction is generated as an adjustment failure result, and the current adjustment instruction as an adjustment failure result is added to the prompt word to obtain a new prompt word. The process returns to the step of inputting the prompt word into the adjustment decision model to obtain the adjustment instruction. If the adjustment is determined to be invalid twice in a row, the safety protection mechanism is triggered and recorded as a failure experience.

[0014] Optionally, after polishing, the following may also be included: Collect image data of the workpiece to be polished after polishing; Based on the polished image data, the surface roughness assessment value, polishing defects, and quality score of each region are determined. Based on the quality scores of each region, the quality score of the workpiece to be polished is calculated, and the quality grade is determined. Based on the surface roughness assessment values ​​and grinding defects of each region, as well as the pre-set grinding targets, the completion status of the grinding target tasks is judged. If not completed, the execution record and the abnormal situation are cross-analyzed to determine the cause of the deviation, and the quality score, quality grade, surface roughness evaluation value, grinding defect situation and grinding target task completion status of the workpiece to be ground are used as the evaluation result. The basic information of the workpiece, the semantic vector of the surface, the structured grinding strategy file, the execution record, and the evaluation result are stored as an experience set in the historical experience case library.

[0015] Optional, also includes: Monitor the number of experience sets in the historical experience case library to determine whether the model fine-tuning conditions have been met. If the target is reached, a preset number of experience sets will be selected from the historical experience case library. The polishing strategy generation model was fine-tuned using the selected experience set; Using a pre-created validation set, we validated the improvement rates of process parameter recommendation accuracy, anomaly cause judgment accuracy, and strategy generation quality score of the fine-tuned grinding strategy generation model. Based on the improvement rate of the recommended process parameters, the improvement rate of the accuracy of the judgment of the cause of the anomaly, and the improvement rate of the quality score of the strategy generation, it is determined whether the fine-tuned grinding strategy generation model passes the evaluation. If approved, the refined polishing strategy generation model will be deployed online. If it fails, it will revert to the previous version of the polishing strategy generation model and record the failure log.

[0016] A curved surface grinding device, comprising: The data acquisition module is used to acquire the 3D point cloud data, visual image data and user language commands of the workpiece to be polished; A digital surface model construction module is used to construct a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data. The vertex parameter calculation module is used to calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model; The region division module is used to divide the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and determine the region type label of each region. The surface semantic vector generation module is used to encode the normal vector, curvature data and region type label of each vertex using a pre-trained lightweight graph neural network, and map them to a high-dimensional semantic embedding space to obtain a structured surface semantic vector. The polishing strategy generation module is used to input a pre-trained polishing strategy generation model based on the visual image data, the user language instructions and the structured surface semantic vector to obtain a structured polishing strategy file, wherein the structured polishing strategy file contains differential process parameters corresponding to each region. The grinding execution module is used to control the robot to grind the workpiece to be ground based on the structured grinding strategy file and generate an execution record. The multimodal data acquisition module is used to acquire six-dimensional force signal data and tactile vibration signals in real time during the polishing process; The robot end effector posture position correction module is used to correct the robot end effector posture position in real time based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex and the structured grinding strategy file. An abnormal situation monitoring module is used to determine abnormal situations based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, and to determine whether it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy based on the abnormal situation. The historical adjustment decision retrieval module is used to search for similar adjustment decision records from a pre-created historical experience case library when it is necessary to call a pre-trained adjustment decision model to adjust the polishing strategy, based on the abnormal situation and the structured surface semantic vector corresponding to the current region. The prompt word determination module is used to determine prompt words based on the six-dimensional force signal data, the tactile vibration signal, the structured surface semantic vector corresponding to the current region, and the similar adjustment decision records; An adjustment instruction generation module is used to input prompt words into the adjustment decision model to obtain adjustment instructions. The adjustment decision model is configured to have the ability to obtain a current state context matrix based on the six-dimensional force signal data and the tactile vibration signal, obtain anomaly cause inference results based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain adjustment instructions based on the anomaly cause inference results and the similar adjustment decision records. The grinding strategy optimization and adjustment module is used to adjust the structured grinding strategy file according to the adjustment instructions to obtain the adjusted structured grinding strategy instructions, and return to execute the steps of controlling the robot to grind the workpiece to be ground based on the structured grinding strategy file until the grinding is completed.

[0017] A surface polishing device includes: a memory and a processor; The memory is used to store programs; The processor is used to execute the program to implement the various steps of the surface polishing method described above.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the surface polishing method as described above.

[0019] As can be seen from the above technical solutions, the surface polishing method, apparatus, equipment, and computer-readable storage medium provided in this application include: acquiring three-dimensional point cloud data, visual image data, and user language commands of the workpiece to be polished; constructing a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data; calculating the normal vector and curvature data of each vertex in the triangular mesh digital surface model; dividing the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and determining the region type label of each region; and using a pre-trained lightweight graph neural network to process the normal vector, curvature data, and region type label of each vertex. The process involves encoding and mapping the data to a high-dimensional semantic embedding space to obtain structured surface semantic vectors. Based on the visual image data, user language commands, and the structured surface semantic vectors, a pre-trained grinding strategy generation model is input to obtain a structured grinding strategy file. This file contains differential process parameters corresponding to each region. Based on the structured grinding strategy file, the robot is controlled to grind the workpiece and an execution record is generated. Six-dimensional force signal data and tactile vibration signals are collected in real time during the grinding process. Based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vectors of each vertex, and the structured grinding strategy file, the process is modified in real time. The robot's end effector is positioned correctly. Based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and a pre-set anomaly threshold, anomalies are identified. Based on these anomalies, it is determined whether a pre-trained adjustment decision model needs to be invoked to adjust the polishing strategy. If so, based on the anomalies and the structured surface semantic vector corresponding to the current region, similar adjustment decision records are searched from a pre-created historical experience case library. Based on the six-dimensional force signal data, the tactile vibration signal, the structured surface semantic vector corresponding to the current region, and the similar adjustment decision records, a prompt word is determined. The prompt word is then input into the adjustment decision... The adjustment decision model is configured to generate adjustment instructions based on the six-dimensional force signal data and the tactile vibration signal, obtain a current state context matrix, obtain anomaly cause inference results based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain adjustment instructions based on the anomaly cause inference results and similar adjustment decision records. Based on the adjustment instructions, the structured grinding strategy file is adjusted accordingly to obtain the adjusted structured grinding strategy instructions, and the process is returned to execute the steps of controlling the robot to grind the workpiece based on the structured grinding strategy file until grinding is completed. This application generates a grinding strategy based on surface semantic features and adjusts the grinding strategy in real time based on the state data during the grinding process to control the robot to achieve surface grinding. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 A flowchart of a surface polishing method provided in this application embodiment; Figure 2 This is a schematic diagram of a curved surface grinding device provided in an embodiment of this application; Figure 3 This is a hardware structure block diagram of a curved surface grinding device provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Figure 1 A flowchart of a surface polishing method provided in this application embodiment may include the following steps: Step S100: Obtain the three-dimensional point cloud data, visual image data, and user language commands of the workpiece to be polished.

[0024] Specifically, to achieve complete coverage of complex curved surfaces, the robot can carry a binocular structured light camera and an RGB industrial camera to perform a global scan of the workpiece to be polished along a preset scanning path. After multi-viewpoint data scanning and stitching, a complete point cloud of the workpiece to be polished is obtained, eliminating blind spots and occlusion problems from a single perspective.

[0025] The specific scanning method can be as follows: a binocular structured light camera projects speckle structured light, and the left and right cameras simultaneously acquire infrared and color images. High-density 3D point clouds are obtained using phase decoding, while surface texture information is acquired using an RGB camera for subsequent semantic-assisted judgment and material recognition. A panoramic scanning imaging mode is adopted, and the complete 3D reconstruction of the workpiece is achieved by fusing color and infrared scene information, binocular vision, and speckle structured light depth information at different locations.

[0026] The binocular structured light 3D reconstruction achieves a reconstruction accuracy of up to 0.02mm, meeting the measurement requirements for high-precision curved surface grinding. During the scanning process, it can output: 3D point cloud data, RGB images, and depth images. The 3D point cloud data includes the spatial coordinates (X, Y, Z) of each point in PLY / PCD format, with a density ≥100 points / cm² and a coordinate accuracy ≤0.02mm. The RGB images are used for material and surface condition identification, with a resolution ≥1920×1080 pixels and a color space of sRGB. The depth images are used to assist in normal vector calculation and hole filling, with a resolution aligned with RGB and a depth accuracy ≤0.05mm.

[0027] Step S101: Based on the three-dimensional point cloud data, construct a triangular mesh digital surface model of the workpiece to be polished.

[0028] Specifically, Step S102: Calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model.

[0029] Specifically, for each vertex in the triangular mesh digital surface model, the normal vector of that point can be calculated by weighted average of the normal vectors of its neighboring triangular surfaces. The specific calculation method can be to calculate the normal vector of the triangular surface associated with each vertex, and then perform weighted summation and normalization using the area of ​​the triangular surface as the weight to calculate the normal vector corresponding to that vertex.

[0030] For each vertex, based on the corresponding normal vector and neighborhood topological relationships, calculate the principal curvature values ​​k1 and k2 for that vertex, and use the principal curvature values ​​to calculate the corresponding Gaussian curvature K and mean curvature H using the following formula: ; .

[0031] The sign of Gaussian curvature can be used to determine the type of surface. For example, K>0 indicates an elliptical point, i.e., a hyperbola region; K=0 indicates a parabolic point, i.e., a simple curvature region; and K<0 indicates a hyperbolic point, i.e., a saddle-shaped region.

[0032] Step S103: Based on the normal vector and curvature data of each vertex, divide the triangular mesh digital surface model into regions and determine the region type label of each region.

[0033] Specifically, when dividing regions, low curvature regions can be prioritized to avoid oversegmentation in abrupt change regions.

[0034] Step S104: Using a pre-trained lightweight graph neural network, the normal vectors, curvature data and region type labels of each vertex are encoded and mapped to a high-dimensional semantic embedding space to obtain structured surface semantic vectors.

[0035] Specifically, lightweight graph neural networks can employ graph neural networks (GNNs). GNNs can process graph structure data composed of grid vertices, capture the neighborhood relationships between vertices, and use discrete Ricci curvature as the core feature for local structure encoding. In graph neural networks, the performance of downstream tasks can be significantly improved, transforming the physical workpiece surface into structured semantic knowledge that can be used for inference by the grinding strategy generation model.

[0036] The normal vectors, curvature data, and region type labels obtained in the preceding steps constitute the geometric feature vector of each vertex. A GNN is used to process these geometric feature vectors, outputting a fixed-dimensional surface semantic vector. This surface semantic vector can include region type (flat, gentle curve, steep curve, abrupt change), processing difficulty (easy, medium, difficult), curvature distribution (uniform, gradual, abrupt change), and recommended processing preference (high speed low force, medium speed medium force, low speed high force, low speed low force multiple passes). Region type can be directly mapped based on region segmentation results. A curvature abrupt change region is defined as a strip-shaped region where the rate of change of the angle between the normal vectors at the boundary of two regions exceeds a set threshold, such as a band-shaped region where the rate of change of the angle between the normal vectors of adjacent points is ≥30° / mm. Processing difficulty can be comprehensively evaluated by considering factors such as the absolute value of curvature, the rate of change of curvature, and the region area. Curvature distribution can be determined by the variance of curvature within that region.

[0037] The resulting surface semantic vector can be structured data in JSON format, which can be directly used as input for subsequent steps. The specific field definitions are as follows: json { "Surface semantic vector": { "Workpiece ID": "W20260710-001", Global Statistics: { Total number of vertices: 245678 Total area: 2.35 "Overall mean curvature": 0.023, "Overall curvature variance": 0.008 }, "Region List": [ { "Region ID": 1, Region Type: Hypercurvature Region "Region boundary vertex index": [1024, 1025, ...], Area: 0.85 Mean curvature: 0.045 "Curvature Distribution": "Gradual Change" "Processing difficulty": "Difficult" Recommended process preference: Low speed, low force, multiple passes. "Recommended Process Parameters": { Recommended target force range: [5.0, 8.0] Recommended speed range: [150, 250] Recommended tool granularity range: ["P120", "P180"] }, "Normal vector distribution": { Mean: [0.12, 0.89, -0.44] Variance: 0.032 } }, { "Region ID": 2, "Region Type": "Single Curvature Region" "Region Boundary Vertex Index": [2048, 2049, ...], Area: 1.20 Mean curvature: 0.012 "Curvature distribution": "Uniform" Processing difficulty: Medium Recommended process preference: Medium speed and medium force "Recommended Process Parameters": { Recommended target force range: [8.0, 12.0] Recommended speed range: [250, 350], Recommended tool granularity range: ["P80", "P120"] }, "Normal vector distribution": { Mean: [0.05, 0.98, -0.19] Variance: 0.005 } }, { "Region ID": 3, Region Type: Region with Abrupt Curvature Change "Region boundary vertex index": [3072, 3073, ...], Area: 0.05 Mean curvature: 0.089 "Curvature Distribution": "Sudden Changes" "Processing difficulty": "Extremely difficult" Recommended process preference: Low speed, low force, multiple passes. "Recommended Process Parameters": { Recommended target force range: [3.0, 5.0] Recommended speed range: [80, 150] Recommended tool granularity range: ["P180", "P240"] }, "Normal vector distribution": { Mean: [0.23, 0.67, -0.71] Variance: 0.089 } } ], "Inter-regional transitional relationships": [ {"Source Region": 1, "Target Region": 2, "Transition Type": "Curvature Gradient", "Transition Band Width": 15.0}, {"Source Region": 2, "Target Region": 3, "Transition Type": "Curvature Abrupt Change", "Transition Band Width": 2.5} ], Semantic encoding confidence: 0.92 } } Step S105: Based on visual image data, user language instructions, and structured surface semantic vectors, input the pre-trained polishing strategy generation model to obtain a structured polishing strategy file.

[0038] Specifically, the structured polishing strategy file contains the differential process parameters corresponding to each region. The polishing strategy generation model does not simply call a preset template, but is based on the "understanding" of the geometric semantics of the surface, performs causal reasoning and dynamic decision-making, and outputs a complete polishing strategy scheme that includes region division, process parameter set and initial path.

[0039] Step S106: Based on the structured grinding strategy file, control the robot to grind the workpiece to be ground and generate an execution record.

[0040] Specifically, the structured polishing strategy file generated in the above steps is converted into real-time motion and force control commands for the robot end effector, driving the polishing tool to complete a high-quality polishing operation along the curved surface and generating an execution record.

[0041] Extract the differentiated process parameters for the current region from the structured grinding strategy file generated in the above steps, initialize the impedance controller parameters, and set the force control mode. Specifically, based on the surface region ID where the current tool is located, retrieve the corresponding differentiated process parameter set from the strategy file to obtain the target normal contact force F. d Feed rate V d Impedance stiffness coefficient K d Impedance damping coefficient B d The parameters are loaded into the control law of the adaptive impedance controller to achieve precise tracking of the contact force. The core model of the impedance control in this scheme is a mass-spring-damped system, and its control equation can be expressed as: ; in, The inertia matrix; For the desired acceleration; This is the actual acceleration; Here is the damping matrix; For the desired speed; This refers to the actual speed; Here is the stiffness matrix; For the desired position; This refers to the actual location; The target normal contact force; This is the actual normal contact force.

[0042] Traditional impedance control uses a fixed stiffness / damping coefficient, which is prone to force tracking deviation or system oscillation in regions with large curvature changes. This scheme adopts adaptive impedance control, where the stiffness / damping coefficient is dynamically generated according to the surface semantics and smoothly updated during regional transitions.

[0043] Step S107: Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process.

[0044] Specifically, a six-dimensional force sensor can be installed between the robot's end flange and the grinding tool to collect six-dimensional force sensor data at a high-frequency sampling rate, and to collect forces (F) in three directions in real time. x ,F y ,F z ) and torque (M) x M y M z The sampling frequency can be set to ≥1kHz. The output of the six-dimensional force sensor is an analog voltage signal, which is converted into a digital signal by an analog-to-digital converter and then transmitted to the main control chip via SPI communication.

[0045] Since the original signal contains high-frequency noise and power frequency interference, a hybrid filtering strategy can be adopted. First, Kalman filtering is performed to remove Gaussian noise, and then weighted sliding filtering is performed to improve signal smoothness. While maintaining a fast response to real force changes, the smoothness of the force signal is significantly improved.

[0046] In addition, the weight of the grinding tool itself may cause a deviation in the force sensor reading. Therefore, the system can perform tool weight calibration before execution to obtain the tool's center of gravity position and weight, and deduct the gravity component in real time during operation to ensure that the sensor output only reflects the grinding contact force.

[0047] Step S108: Based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex and the structured grinding strategy file, the robot end effector position is corrected in real time.

[0048] Specifically, when the surface geometry is unknown or the curvature changes continuously, the tool's posture is aligned with the surface normal in real time.

[0049] Step S109: Based on the six-dimensional force signal data, tactile vibration signal, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, determine the abnormal situation, and based on the abnormal situation, determine whether it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy.

[0050] Specifically, the force control status is monitored in real time to detect contact force deviation, abnormal vibration, tool overload, and other conditions, and a state semantic vector is generated. If it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy, step S110 is executed; if it is not necessary to call the pre-trained adjustment decision model to adjust the polishing strategy, the original polishing strategy is continued.

[0051] Unlike the offline / quasi-online strategy generation that produces structured refinement strategy files in the previous steps, strategy adjustment needs to be performed in real-time online mode, requiring inference latency ≤100ms. To balance response speed and decision quality under limited computing power, the following engineering optimization strategies can be adopted: Cascaded decision architecture: High-frequency anomaly detection is handled by a lightweight rule / threshold module (response time <10ms), and the large model is only triggered for deep inference when an anomaly is detected or a significant change in state occurs (response time <100ms), avoiding continuous high-load inference. Model distillation and quantization: The large model is distilled into lightweight small models and deployed at the edge to handle only routine inference tasks; for complex anomaly scenarios, key data is uploaded to the large model in the cloud for deep analysis through an edge-cloud collaboration mechanism. Context caching and sparse inference: The inference results of previous similar states are reused, and only newly emerging feature differences are input into the model to reduce redundant calculations. Through the above strategies, the real-time constraints can be met while ensuring decision intelligence.

[0052] Step S110: Based on the abnormal situation and the structured surface semantic vector corresponding to the current region, search for similar adjustment decision records from the pre-created historical experience case library.

[0053] Specifically, similar adjustment decision records are searched from a pre-created historical experience case library, ensuring that the adjustment strategy is not generated out of thin air, but rather selected and parameterized from a predefined strategy template library. These adjustment strategy templates can originate from: process rules from domain experts, generalizations from historical successful cases, and pre-calculations of physical simulation models.

[0054] Step S111: Based on the six-dimensional force signal data, tactile vibration signal, structured surface semantic vector corresponding to the current area, and similar adjustment decision records, determine the prompt words.

[0055] Specifically, six-dimensional force signal data, tactile vibration signals, the structured surface semantic vector corresponding to the current region, and similar adjustment decision records are assembled into structured prompt words and input into the large model. An example of a prompt word template is shown below: [Task] Analyze the causes of the current abnormal polishing status and generate adjustment strategies. [Current Region Semantics] - Region type: Hypercurvature region - Processing difficulty: High - Curvature distribution: Gradual - Expected target force: 6.0N [Current State Context] - Actual normal force: 7.2N (deviation +1.2N, trend: continuously increasing) - Force fluctuation variance: 0.32 N² (normal range) - Vibration frequency: 1350Hz (normal range) - Acoustic emission RMS: 0.45 (close to the warning threshold) - Current grinding stage: 2nd pass (out of 2) [Historical Experience Reference] - There are cases where, under similar working conditions, the increased curvature of CFRP leads to a reduction in contact area. [Reasoning Requirements] 1. Determine the cause of the anomaly (select or combine from multiple candidate causes) 2. Assess the confidence level of the causal factors. 3. Generate specific adjustment instructions Step S112: Input the prompt words into the adjustment decision model to obtain the adjustment instructions.

[0056] Specifically, the adjustment decision model is configured to obtain a current state context matrix based on six-dimensional force signal data and tactile vibration signals; obtain anomaly cause inference results based on the current state context matrix and the structured surface semantic vector corresponding to the current region; and obtain adjustment instructions based on the anomaly cause inference results and similar adjustment decision records. The core task of the adjustment decision model is to select the correct strategy template from similar adjustment decision records and fill in specific parameter values ​​according to the current state. This ensures both the controllability of the large model output and retains its advantage of flexible reasoning.

[0057] The decision-making model is adjusted to select and combine the following candidate causes: surface geometry factors: changes in contact area due to increased / decreased curvature, and attitude tracking deviation due to abrupt changes in the surface normal vector; tool condition factors: grinding efficiency decreases due to belt passivation / wear, and improper belt grit selection; workpiece material factors: uneven material hardness, foreign matter / impurities on the surface, and thermal deformation due to differences in material thermal conductivity; process parameter factors: improper target force setting, and mismatch between feed rate and surface curvature; system factors: zero drift of the force sensor and accumulation of gravity compensation error.

[0058] Adjusting the output of the decision model to a structured reasoning result can include: primary cause analysis, secondary cause analysis, and confidence assessment.

[0059] The decision-making model can be adjusted based on the causal reasoning results, selecting one or more of the following adjustment strategies in combination: force control parameter correction, adjusting the target force Fd, stiffness coefficient Kd, and damping coefficient Bd, suitable for scenarios with curvature changes and unstable contact; dynamic parameter adjustment, reducing / increasing the feed rate Vd and slow restarting after pausing, suitable for scenarios with sudden curvature changes and vibration risks; path replanning, generating locally corrected paths to bypass abnormal areas, suitable for scenarios with workpiece surface defects and hard points; tool replacement suggestion, prompting "recommend replacing the sanding belt" and adjusting subsequent parameter compensation, suitable for scenarios with sanding belt passivation / wear; task termination / pause, issuing an emergency pause command, suitable for scenarios with severe anomalies and safety risks.

[0060] Step S113: Based on the adjustment instructions, adjust the structured polishing strategy file accordingly to obtain the adjusted structured polishing strategy instructions.

[0061] Specifically, after receiving the adjusted structured grinding strategy instructions, the system returns to execute the steps based on the structured grinding strategy file, controlling the robot to grind the workpiece until grinding is complete. The adjustment instructions are mapped to control parameter update instructions, the format of which can be as follows: json { Command ID: "ADJ-20260710-143522-001", "Command Type": "Force Control Parameter Correction", "Target Module": "Adaptive Impedance Control Submodule", "Parameter Update": { "F_d": 5.0, "K_d": 1500, "B_d": 0.8, Gradient method: "S-curve" "Gradual Time": 200 }, Execution priority: "high", "Validity period": "Until the next regional switch", "Large Model Inference Confidence": 0.92 } Adjusting decisions requires extremely strict real-time performance. To balance real-time performance with decision intelligence, a three-tiered cascaded architecture can be adopted: Tier 1 (high frequency, <10ms): rule-based threshold detection filters out most normal states, only propagating events exceeding the threshold; Tier 2 (medium frequency, 10~50ms): a lightweight machine learning classifier (such as a random forest) quickly classifies warning-level events; Tier 3 (low frequency, <100ms): deep inference using a large model handles only complex anomaly scenarios that Tier 2 cannot diagnose. Furthermore, model distillation compresses the large model into a lightweight version deployable at the edge. For extremely complex anomaly scenarios, a collaborative edge-cloud mechanism uploads key data to the large model in the cloud for in-depth analysis.

[0062] This application provides a surface polishing method, comprising: acquiring three-dimensional point cloud data, visual image data, and user language commands of the workpiece to be polished; constructing a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data; calculating the normal vector and curvature data of each vertex in the triangular mesh digital surface model; dividing the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and determining the region type label of each region; and encoding the normal vector, curvature data, and region type label of each vertex using a pre-trained lightweight graph neural network, mapping them to a high-dimensional semantic embedding space. The process involves obtaining structured surface semantic vectors; inputting these vectors into a pre-trained grinding strategy generation model based on visual image data, user language commands, and the structured surface semantic vectors, resulting in a structured grinding strategy file containing differential process parameters for each region; controlling the robot to grind the workpiece based on the structured grinding strategy file and generating execution records; real-time acquisition of six-dimensional force signal data and tactile vibration signals during the grinding process; and real-time correction of the robot's end-effector pose based on the six-dimensional force signal data, differential process parameters for each region, normal vectors of each vertex, and the structured grinding strategy file. The system identifies the state location; based on six-dimensional force signal data, tactile vibration signals, differential process parameters corresponding to each region, and pre-set anomaly thresholds, it determines abnormal situations and, based on these abnormal situations, decides whether to call a pre-trained adjustment decision model to adjust the polishing strategy; if so, it searches for similar adjustment decision records from a pre-created historical experience case library based on the abnormal situation and the structured surface semantic vector corresponding to the current region; based on six-dimensional force signal data, tactile vibration signals, the structured surface semantic vector corresponding to the current region, and similar adjustment decision records, it determines prompt words; and inputs the prompt words into the adjustment decision model. Upon receiving adjustment instructions, the adjustment decision model is configured to obtain a current state context matrix based on six-dimensional force signal data and tactile vibration signals. Based on the current state context matrix and the structured surface semantic vector corresponding to the current region, it obtains anomaly cause inference results. Based on the anomaly cause inference results and similar adjustment decision records, it obtains adjustment instructions. Based on these instructions, the structured grinding strategy file is adjusted accordingly to obtain the adjusted structured grinding strategy instructions. The system then returns to execute the steps based on the structured grinding strategy file, controlling the robot to grind the workpiece until grinding is complete. This application generates a grinding strategy based on surface semantic features and adjusts the grinding strategy in real time based on the state data during the grinding process, controlling the robot to achieve surface grinding.

[0063] In some embodiments of this application, the acquired raw 3D point cloud contains noise and outliers, and the data volume is huge, requiring multi-level preprocessing. Therefore, step S101, based on the 3D point cloud data, constructs a triangular mesh digital surface model of the workpiece to be polished, including: S11. For each point, calculate the average distance from all points in its neighborhood to that point.

[0064] S12. Points whose average distance exceeds the preset global average distance threshold are identified as outliers and removed.

[0065] Specifically, the preset global average distance threshold can be set to 2-3 times the standard deviation.

[0066] S13. Divide the three-dimensional point cloud space into a voxel grid of a specified size. With the standard of retaining only one representative point in each grid, simplify the three-dimensional point cloud data after removing outlier noise to obtain the simplified three-dimensional point cloud data.

[0067] Specifically, to balance accuracy and computational efficiency, voxel grid downsampling can be used to divide the point cloud space into voxel grids of a specified size (such as 1mm×1mm×1mm). Only one representative point is retained in each grid, usually the centroid or the point closest to the centroid, thereby significantly compressing the amount of data while preserving the geometric features of the surface.

[0068] S14. Based on the simplified 3D point cloud data, construct a continuous triangular mesh digital surface model of the workpiece to be polished.

[0069] Specifically, the simplified scattered 3D point cloud data is reconstructed into a continuous triangular mesh digital surface model. By implicitly fitting a global indicator function, a watertight and smooth curved surface mesh can be generated, which facilitates subsequent curvature calculation.

[0070] In some embodiments of this application, the curvature data includes average curvature. Step S103 involves dividing the triangular mesh digital surface model into regions based on the normal vectors and curvature data of each vertex, and determining the region type label for each region, including: S21. From the vertices whose region type labels have not been determined, select the vertex corresponding to the minimum average curvature as the current seed point, and determine the region type label of the region where the current seed point is located based on the normal vector and curvature data of the current seed point.

[0071] Specifically, selecting the point with the smallest curvature as the seed point allows for growth from low-curvature regions to high-curvature regions, ensuring that flat regions are segmented first, thereby avoiding over-segmentation in abrupt regions.

[0072] The region type label of the current seed point's location can be determined based on the Gaussian curvature and average curvature in the normal vector and curvature data. For example, the region type label of a seed point with Gaussian curvature K≈0 and average curvature H≈0 is a planar region, and the grinding process semantics can be understood as easy to process, allowing for high-speed and high-pressure grinding; the region type label of a seed point with Gaussian curvature K≈0 and average curvature H≠0 is a single curvature region, such as a cylindrical or conical surface, and the grinding process semantics can be understood as medium difficulty, requiring maintaining the tool posture; the region type label of a seed point with Gaussian curvature K>0 is a double curvature region, such as a sphere or freeform surface, and the grinding process semantics can be understood as difficult, requiring variable force and speed grinding; the region type label of a seed point where the rate of change of the angle between the normal vectors of adjacent vertices exceeds a preset abrupt change threshold is a curvature abrupt change region, and the grinding process semantics can be understood as extremely difficult, requiring slowing down and depressurizing for fine processing.

[0073] S22. Incorporate the neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold into the region where the current seed point is located.

[0074] Specifically, the preset threshold for the angle between the normal vectors can be adjusted according to the complexity of the surface. If the surface is complex, the threshold is reduced to achieve finer segmentation; if the surface is simple, the threshold is increased to achieve coarser segmentation. The default value can be set to 15°.

[0075] S23. Determine if there are any newly added neighboring vertices in the region where the current seed point is located.

[0076] Specifically, if there are neighboring vertices newly included in the region where the current seed point is located, then execute S24; if there are no neighboring vertices newly included in the region where the current seed point is located, then return to execute S21, the step of selecting the vertex with the minimum average curvature from the vertices whose region type labels have not been determined as the current seed point.

[0077] S24. The neighboring vertices newly included in the region of the current seed point are sequentially used as new current seed points, and the process is returned to execute the step of including the neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold into the region of the current seed point.

[0078] Specifically, the neighboring vertices of the region where the newly included current seed point is located are successively used as new current seed points, and the region is further explored in the surrounding area. The region is divided by using the region growing method.

[0079] In some embodiments of this application, the polishing strategy generation model includes: an input layer, a multimodal semantic fusion layer, a polishing state understanding and process reasoning layer, a partitioning strategy and process parameter generation layer, a baseline polishing path generation layer, and a structured polishing strategy file output layer.

[0080] S31. Obtain visual image data, user language commands, and structured surface semantic vectors through the input layer.

[0081] Specifically, user language commands can include processing objectives, materials, quality requirements, thickness of material to be removed, and efficiency priorities (quality priority / efficiency priority / balance) input by the user through natural language. For example, for surface roughening treatment of carbon fiber reinforced composite polymer skin, the target roughness Ra≤1.6μm can be achieved by performing intent recognition and entity extraction on user language commands, and key information can be extracted: workpiece material (carbon fiber reinforced composite polymer), processing objective (roughening), and quality index (Ra≤1.6μm).

[0082] S32. Through the multimodal semantic fusion layer, the input visual image data, user language commands, and structured surface semantic vectors are mapped to a unified semantic embedding space to achieve adaptive fusion of cross-modal features and obtain the fused semantic features.

[0083] Specifically, a multimodal unified representation technique is adopted to map structured surface semantic vectors, visual features, and language instructions to a unified semantic embedding space. Learnable attention weights are introduced between the visual encoder and the text encoder to achieve adaptive fusion of cross-modal features and obtain fused semantic features.

[0084] S33. Through the grinding state understanding and process reasoning layer, based on the fused semantic features, determine the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each area, and the current grinding stage.

[0085] Specifically, the material identification and characteristic retrieval of the workpiece to be polished can be performed by identifying the workpiece material from the language instructions and visual texture features in the fused semantic features, such as carbon fiber reinforced composite polymer, aluminum alloy, and atomic putty. The physical characteristics of the material can be retrieved from the built-in domain knowledge graph: hardness level, such as carbon fiber reinforced composite polymer has high hardness, poor thermal conductivity, and is sensitive to overheating; recommended abrasive type, such as ceramic abrasive is recommended for carbon fiber reinforced composite polymer; force-sensitive characteristics, such as the putty layer is prone to exposing the substrate if over-polished.

[0086] A comprehensive assessment of the overall processing difficulty level of each region can be conducted by combining the processing difficulty field in the surface semantic vector with the characteristics of the workpiece material identified in language instructions and visual texture features. Each region can be classified into low, medium, high, and very high difficulty levels. For example, by utilizing region type (hypercurvature region), curvature distribution (gradient), and material (carbon fiber reinforced composite polymer), it can be inferred that the region has high curvature and the material is difficult to process, thus determining that the overall processing difficulty level of this region is high.

[0087] The determination of the current grinding stage can be based on the machining target and the current surface condition in the user's language instructions to determine the main grinding stage of this task, providing a benchmark for subsequent parameter generation. The machining target can be rough grinding, semi-finish grinding, or finish grinding.

[0088] S34. Through the partitioning strategy and process parameter generation layer, based on the comprehensive processing difficulty level of each region and the spatial adjacency relationship between each region, the processing order between regions is generated, and based on the processing order between regions and the fused semantic features, the differentiated process parameters corresponding to each region are generated.

[0089] Specifically, based on the spatial adjacency relationship and processing difficulty of each region, the optimal processing sequence is generated. Specific principles may include: processing regions with lower processing difficulty first, and processing regions with higher processing difficulty later; avoiding frequent switching of tools between regions with different curvatures; and considering robot accessibility constraints.

[0090] For each region, the material hardness, region curvature, grinding stage, region area, processing difficulty, target efficiency, material properties, target roughness, surface stiffness, contact stability requirements, system stability, response speed requirements, surface normal vector distribution, processing difficulty, and allowance size can be determined based on the fused semantic features.

[0091] Determine the target normal contact force F by utilizing material hardness, regional curvature, and grinding stage. d For example, 5~12N; determine the recommended tool grit size based on area, processing difficulty, and target efficiency, such as 150~350mm / s; determine the recommended tool grit size and sandpaper mesh number based on material properties and target roughness, such as P80~P240; determine the impedance control stiffness coefficient K based on surface stiffness and contact stability requirements. d For example, 1200~3000 N / m; determine the impedance control damping coefficient B based on system stability and response speed requirements. dThe tool posture constraints, such as the normal tracking angle deviation, are determined using the surface normal vector distribution. The number of tool passes is determined using the machining difficulty and allowance size, and the number of tool passes is an integer, such as 1 to 3. The parameter generation can be implemented using a retrieval-enhanced generation mechanism, which uses the current region features as query conditions to retrieve similar cases from a pre-built process parameter case library, and then fine-tunes the output based on the current specific features.

[0092] Because abrupt changes in process parameters between different regions can cause uneven surface quality, a smooth transition planning of parameters between regions can be achieved based on the processing sequence between regions. By using gradual force control curves and speed transition curves, linear or S-shaped gradual changes of parameters can be achieved within the width of the transition zone. Through collaborative learning, continuous mapping of parameters at the region boundaries can be achieved, and finally, the differentiated process parameters corresponding to each region can be determined.

[0093] S35. Through the reference polishing path generation layer, for each region, based on the region type label corresponding to each region, different path generation strategies are adopted to create the target polishing points corresponding to each region. According to the pre-set objective function, the smoothness of the target polishing points is optimized, and the polishing path corresponding to each region is determined. Based on the processing order between regions and the polishing path corresponding to each region, the transition relationship between regions is determined.

[0094] Specifically, the optimal path generation strategy can be automatically matched according to the region type. For example, the isoparametric method can be used for planar regions or single-curvature regions to generate paths along the surface parameter lines, which is computationally efficient and simple to implement. The isoparametric method can be used for double-curvature regions to ensure that the residual height between each path is consistent and to obtain uniform surface quality. The adaptive densification path can be used for curvature abrupt change regions to increase the path density within the abrupt change zone and ensure that the tool covers the area sufficiently.

[0095] For each region, a sufficient number of target grinding points are created based on a defined path generation strategy to ensure that the path fully meets the processing requirements. Each target grinding point may include: spatial coordinates (X, Y, Z), tool posture (normal vector alignment), desired contact force, and desired speed.

[0096] To optimize the smoothness of the path in the joint space, an objective function can be pre-defined for optimization. The objective function takes into account the following factors: shortest path length, minimized joint impact, and balanced tool wear.

[0097] After the paths for each region are generated, the transitional connection paths between regions are planned based on the processing order between regions to ensure that the tool can smoothly move from the end of the current region to the starting point of the next region, avoiding collisions and singularities.

[0098] S36. Through the structured grinding strategy file output layer, a structured grinding strategy file is generated based on the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each region, the current grinding stage, the differentiated process parameters corresponding to each region, the path generation strategy, the target grinding point number corresponding to each region, and the transition relationship between regions.

[0099] Specifically, the structured polishing strategy file can be output in JSON format, and the specific field definitions can be as follows: json { "Strategy ID": "POLICY-20260710-001", "Generated timestamp": "2026-07-10T14:32:18Z", "Large Model Version": "GrindLLM-v2.3.1", Strategy Summary: { Total number of processing areas: 3 Estimated processing time: 420, Recommended tool: Ceramic abrasive belt, grit size P120-P180 }, "Regional Processing Plan": [ { "Region ID": 1, "Area Name": "Hypercurvature Head Area", Processing difficulty: "High" Grinding Stage: Semi-finish Grinding "Process parameters": { "Target normal contact force_Fd": {"Value": 6.0, "Unit": "N", "Allowable deviation": "+ / -0.5"}, "Feed rate_Vd": {"Value": 200, "Unit": "mm / s", "Tolerance": "+ / -10%"}, Recommended tool granularity: P150 "Impedance stiffness_Kd": 1500, "Impedance Damping_Bd": 0.8, Number of cuts: 2 "Adjacent path overlap rate": 0.3 }, "Path Strategy": "Equal Residual Height Method", "Number of critical points on the path": 385 "Transitional Relationship": { "Adjacent area": ​​2, Transition band width: 12.0 "Parameter Gradient Method": "S-curve" } }, { "Region ID": 2, "Area Name": "Middle section of single-curvature fuselage", Processing difficulty: Medium Grinding Stage: Semi-finish Grinding "Process parameters": { "Target normal contact force_Fd": {"Value": 10.0, "Unit": "N", "Allowable deviation": "+ / -0.8"}, "Feed rate_Vd": {"Value": 300, "Unit": "mm / s", "Tolerance": "+ / -10%"}, Recommended tool granularity: "P120" "Impedance stiffness_Kd": 2200, "Impedance Damping_Bd": 0.6, Number of cuts: 1 "Adjacent path overlap rate": 0.2 }, "Path Strategy": "Equal Parameter Method" "Number of critical points on the path": 520, "Transitional Relationship": { "Adjacent region": [1, 3], "Transition band width": [12.0, 5.0], "Parameter Gradient Method": "S-curve" } }, { "Region ID": 3, "Region Name": "Wing Connection Abrupt Change Region", Processing difficulty: "Extremely high" Grinding Stage: Semi-finish Grinding "Process parameters": { "Target normal contact force_Fd": {"Value": 4.0, "Unit": "N", "Allowable deviation": "+ / -0.3"}, "Feed rate_Vd": {"Value": 120, "Unit": "mm / s", "Tolerance": "+ / -10%"}, Recommended tool granularity: P200 "Impedance stiffness_Kd": 1000, "Impedance Damping_Bd": 0.9, Number of cuts: 3 "Adjacent path overlap rate": 0.4, "Path-adaptive encryption multiplier": 2.0 }, "Path Strategy": "Adaptive encryption and other residual height methods", "Number of critical points on the path": 156 "Transitional Relationship": { "Adjacent area": ​​2, Transition zone width: 3.0 "Parameter Gradient Method": "Linear Gradient" } } ], "Global path metadata": { Total path length: 12500 Total critical points: 1061 Estimated processing time: 420, Starting point coordinates: [125.6, 340.2, 85.0] "Termination point coordinates": [680.3, 420.8, 92.5] }, Policy confidence level: 0.89 } Furthermore, the construction process of the polishing strategy generation model relies on a high-quality domain knowledge base. The knowledge base construction involves the following stages: multi-source data acquisition: extracting knowledge from academic papers, process manuals, patent documents, and expert experience; task-oriented preprocessing: cleaning and reconstructing to retain information closely related to process parameters; domain knowledge tag definition: constructing an entity-relationship-attribute tag system; knowledge-driven data augmentation: expanding sample diversity based on physical constraints. To ensure the accuracy and response speed of inference, the polishing strategy generation model can adopt a two-stage hybrid fine-tuning approach: the first stage is full-parameter fine-tuning: based on a general multimodal large model, full-parameter fine-tuning is performed using polishing domain process data to improve domain adaptability; the second stage is efficient parameter fine-tuning: using a lightweight method, incremental fine-tuning is performed for specific workpiece types, updating only a small number of parameters.

[0100] To ensure that the real-time performance of the polishing process is not affected, the policy generation step can be performed offline or quasi-online. Policy generation is completed before the polishing process begins, without consuming real-time computing resources during the polishing process. The target latency for a single policy generation inference is ≤3 seconds. Model quantization compression and sparse attention techniques are used to reduce computational overhead.

[0101] In some embodiments of this application, the differential process parameters include the target normal contact force. Step S108, based on six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex, and the structured grinding strategy file, real-time correction of the robot's end effector posture position includes: S41. Determine the normal vector of the current contact point based on the normal vectors of the three vertices on the triangular surface where the current contact point is located.

[0102] Specifically, during the polishing process, the surface normal vector corresponding to the robot's current position can be obtained in real time by interpolation from the triangular mesh digital surface model constructed in the above steps. By locating the triangular surface patch to which the current tool contact point belongs in the mesh, the normal vectors of the three vertices are interpolated using the centroid coordinates to obtain the estimated value of the continuous normal vector of the current contact point.

[0103] S42. Based on the six-dimensional force signal data and normal vector of the current contact point, determine the current normal contact force.

[0104] Specifically, since the grinding contact force is along the normal direction of the curved surface, the normal component needs to be extracted from the six-dimensional force. This can be achieved by taking the dot product of the force vector measured by the six-dimensional force sensor and the normal vector of the current contact point to obtain the normal contact force F. n : ; in, The force vector measured by the six-dimensional force sensor; This is the normal vector of the current contact point.

[0105] S43. Calculate the deviation between the current normal contact force and the target normal contact force in real time.

[0106] Specifically, the current normal contact force F is calculated in real time. n Contact force F with the target normal direction d deviation value F: .

[0107] S44. Using the deviation value and the differential process parameters corresponding to the current area, calculate the robot end position correction amount, and based on the robot end position correction amount, correct the robot end position in the normal direction of the current contact point.

[0108] Specifically, according to the impedance control equation, the force deviation F is mapped to the end position correction amount. x. In the discrete control cycle, the update law for the position correction can be: ; in, This is the adjustment amount for the end position of the next control cycle (step k+1); This is the correction amount for the end position of the current control cycle (step k); The inertia matrix; This represents the deviation between the current normal contact force and the target normal contact force. Here is the damping matrix; For speed error; Here is the stiffness matrix; This refers to the positional error; To control the cycle.

[0109] The position correction is superimposed on the position command of the reference path to generate the final position command, which is then sent to the robot's underlying servo system. To ensure that position control and attitude control do not interfere with each other, an orthogonal decomposition method can be used to decompose the end effector's Cartesian space into force control directions and position control directions. The position control direction is aligned with the surface normal, while the attitude control direction lies within the surface tangent plane. This is achieved through the feature matrix. and Achieve decoupling between the two subspaces: ; ; The robot end-effector position correction is applied to the normal direction of the current contact point. The robot's end-effector position is corrected in space. The robot end-effector attitude correction command acts in the direction of the tangent plane of the surface where the current contact point is located. (Space) to correct the robot's end-effector posture When the tool transitions from one surface region to another, a parameter update command can be sent. The impedance controller receives the new M... d B d K d The parameters are gradually changed within the transition band width using a parameter smoothing transition algorithm, such as an S-curve or linear gradient, to avoid shocks caused by abrupt parameter changes.

[0110] S45. Determine the path tangent direction of the current contact point based on the structured polishing strategy file.

[0111] S46. Based on the normal vector of the current contact point and the tangent direction of the path, determine the desired posture of the robot's end effector.

[0112] Specifically, during the polishing process, the desired posture of the end effector of the robot arm can be set as follows: the tool axis (Z-axis) is aligned with the normal vector of the current contact point, and the tool's forward direction is consistent with the tangent direction of the path.

[0113] S47. Obtain the current posture data of the robot's end effector and calculate the current posture error.

[0114] Specifically, the error between the current attitude and the desired attitude can be described using rotation matrix difference or axis-angle representation.

[0115] S48. Based on the current attitude error and angular velocity error, generate robot end-effector attitude correction instructions, and correct the robot end-effector attitude in the tangent plane direction of the surface where the current contact point is located based on the robot end-effector attitude correction instructions.

[0116] Specifically, based on the attitude error, an attitude correction command can be generated using an attitude controller based on modified Rodrigues parameters (MRP). This method avoids the singularity of Euler angles and the "unwinding" problem of quaternions, and is suitable for a wide range of attitude adjustments. The attitude control law can be: ; in, The desired angular velocity command for the robot's end effector; To control the gain; The MRP representation of attitude error; To control the gain; This represents the attitude angular velocity error at the current moment.

[0117] Attitude angular velocity error ω at the current moment e This refers to the deviation between the desired angular velocity and the actual angular velocity. Desired angular velocity ω d Differential kinematics can be used to calculate the angular velocity required at the current path point to maintain alignment with the normal, which is the desired angular velocity, i.e., the rate of change of the surface normal vector direction; the actual angular velocity ω... actual The data can be retrieved in real time through the robot's motion control system.

[0118] Using the desired angular velocity ω d and actual angular velocity ω actual The attitude angular velocity error ω at the current moment can be calculated. e =ω d ω actual This difference represents the deviation between how fast the tool should rotate and how fast it actually rotates, and is the feedback error term in the attitude control closed loop.

[0119] The robot's end effector posture is adjusted in real time based on the surface normal vector to keep the grinding tool axis coincident with the surface normal and the grinding disc surface in contact with the local tangent plane of the surface.

[0120] In some embodiments of this application, step S109, determining abnormal situations based on six-dimensional force signal data, tactile vibration signals, differential process parameters corresponding to each region, and a pre-set abnormal threshold, includes: S51. Generate real-time force perception semantic vectors based on six-dimensional force signal data.

[0121] Specifically, in each control cycle, the following status indicators are recorded: normal contact force F n Contact force F with the target normal direction d Deviation value and deviation trend, tangential force F t Force signal fluctuation variance and vibration signal spectral characteristics.

[0122] When the following conditions are detected, a force perception semantic vector can be generated: force deviation exceeds a preset dynamic threshold (e.g., ±2N for more than 100ms), force signal fluctuation variance exceeds a threshold (indicating tremor risk), tangential force increases abnormally (indicating tool lag or surface abrupt change), and the force perception semantic vectors of the most recent N cycles are accumulated and stored, which can be used to capture the trend of state changes.

[0123] The format of the force-sensory semantic vector is similar to that of the surface semantic vector, containing structured fields such as deviation type, deviation amount, and deviation trend, which are used by the model to infer the causes of anomalies. The specific content of the force-sensory semantic vector can be shown below: json { "Timestamp": "2026-07-10T14:35:22.145Z", "Current Region ID": 1, "Normal contact force": { Actual value: 7.2 Target value: 6.0 Deviation: +1.2 "Trend of Deviation": "Increasing" }, "Tangential force": { "Fx": 1.8, "Fy": -0.5, "Combined Force": 1.87 }, "Contact stability": { Force fluctuation variance: 0.32, Stability Rating: "Normal" }, "Vibration state": { Clock speed: 1250 Amplitude: 0.18 Risk Indication: No abnormalities }, Status Label: "Force Deviation Positive Deviation_Stable and Controllable" Confidence level: 0.95 } S52. Based on the tactile vibration signal, the vibration feature vector is extracted.

[0124] Specifically, the tactile vibration signal can include acceleration signals acquired at a sampling frequency of ≥2kHz and acoustic emission signals acquired at a sampling frequency of ≥1MHz. Real-time feature extraction is performed at the edge: time domain features: root mean square (RMS), peak value, kurtosis; frequency domain features: extracting the main frequency component and its amplitude through FFT; feature change rate: the trend of feature change between adjacent windows, as the vibration feature vector.

[0125] S53. Using the force perception semantic vector with a preset period and the corresponding vibration feature vector, concatenate them according to the time sequence to construct a state context matrix.

[0126] Specifically, the force perception semantic vector and vibration feature vector are concatenated according to time series to construct a state context matrix: ; in, Let be the force perception semantic vector at time t; Let t be the vibration feature vector at time t. This matrix serves as the input context for the model, enabling the model to see the process of state change rather than just the current instantaneous value.

[0127] S54. Based on the state context matrix, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, perform multi-dimensional abnormal detection to determine the abnormal situation.

[0128] Specifically, the state context matrix includes the actual normal contact force, actual vibration amplitude, actual root mean square value of acoustic emission, actual force signal variance, and actual grinding area. Differential process parameters include the target normal contact force. Based on the state context matrix, the differential process parameters corresponding to each area, and pre-set anomaly thresholds, multi-dimensional anomaly detection is performed to determine anomalies. This can include: if the difference between the actual normal contact force and the target normal contact force is greater than a preset normal contact force anomaly threshold, and the duration is greater than a preset deviation time threshold, then a force deviation warning is issued; if the actual normal contact force... If the rate of change of the difference between the contact force and the target normal contact force is greater than the preset threshold for the rate of change of the normal contact force deviation, a force mutation warning is issued; if the actual vibration amplitude is greater than the preset vibration amplitude threshold, a vibration risk warning is issued; if the actual root mean square value of acoustic emission is greater than the preset root mean square value threshold for acoustic emission, a tool dulling or burn warning is issued; if the actual force signal variance is greater than the preset force signal variance, a contact instability warning is issued; if the actual grinding area is different from the target grinding area, a parameter update warning is issued; the warning situation and the number of warnings are recorded as abnormal situations.

[0129] Furthermore, determining whether to invoke the pre-trained adjustment decision model to adjust the polishing strategy can be based on abnormal situations, specifically divided into three levels: If all indicators have no warnings, it is the normal level, and the polishing strategy remains unchanged, without invoking the pre-trained adjustment decision model to adjust the polishing strategy; if 1-2 indicators exceed the threshold but the deviation is small, it is the warning level, and the adjustment decision model can be invoked for predictive adjustment; if multiple indicators severely exceed the threshold, it is the abnormal level, and the pre-trained adjustment decision model needs to be invoked immediately to adjust the polishing strategy.

[0130] Furthermore, in addition to the three levels, conditions can be added to trigger the adjustment of the decision-making model, such as a continuous deterioration of the force deviation trend (the deviation value increases for three consecutive time windows), the first control cycle after the completion of the regional switch (predictive parameter fine-tuning), and the cumulative number of anomalies exceeding the threshold (such as triggering ≥5 warnings within 10 minutes).

[0131] In some embodiments of this application, after receiving the adjusted structured grinding strategy instructions and controlling the robot to grind the workpiece based on the adjusted structured grinding strategy file, the method further includes: S61. Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process.

[0132] S62. Based on six-dimensional force signal data and tactile vibration signals, monitor the changing trends of abnormal situations.

[0133] Specifically, within the monitoring window after the adjustment command is issued, such as 500ms to 2s, the changing trend of the status indicators is continuously monitored.

[0134] S63. If the abnormal indicators corresponding to the abnormal situation gradually converge to the target value, the adjustment is deemed effective and recorded as a successful experience.

[0135] S64. If the abnormal indicator corresponding to the abnormal situation does not gradually converge to the target value or a new abnormal situation occurs, the adjustment is determined to be invalid, the current adjustment instruction is generated as an adjustment failure result, and the current adjustment instruction as an adjustment failure result is added to the prompt word to obtain a new prompt word. The process of returning to the execution step of inputting the prompt word into the adjustment decision model to obtain the adjustment instruction is completed. If the adjustment is determined to be invalid twice in a row, the safety protection mechanism is triggered and recorded as a failure experience.

[0136] Specifically, if the deviation continues to worsen or new anomalies occur, the adjustment is deemed ineffective, triggering a new round of model inference and employing recursive adjustment; if two consecutive adjustments are ineffective, the safety protection mechanism is triggered, suspending the task and issuing an alarm.

[0137] After each adjustment is completed, the following information will be recorded in the adjustment log: triggering reason, adjustment strategy, execution result (success / partial success / failure), comparison of state before and after adjustment, and complete context of model inference input and output.

[0138] In some embodiments of this application, in order to continuously improve the model's process decision-making capability over time, the following may be included after polishing: S71. Collect image data of the workpiece after grinding.

[0139] Specifically, after polishing, high-resolution images of the workpiece surface are acquired using a visual perception module to provide a data foundation for subsequent quality assessment. An RGB industrial camera is used to acquire multi-view images of the polished workpiece surface under uniform lighting conditions. To ensure the comparability of the assessment results, the acquisition conditions should be consistent with the global scan in step S100. For large workpieces, a robot is used to drive the camera along a preset path to perform scanning images, covering all polished areas.

[0140] Furthermore, the acquired raw images can be preprocessed: by white balance and brightness normalization, the influence of ambient light changes on the evaluation can be eliminated; images from different perspectives can be stitched together to form a complete panoramic view of the workpiece surface; and based on the region division results, image subsets of each curved surface region can be extracted to facilitate regional evaluation.

[0141] S72. Based on the image data after polishing, determine the surface roughness assessment value, polishing defects, and quality score for each area.

[0142] Specifically, this application can use a multi-level convolutional neural network model for roughness evaluation. The model structure can be based on the ResNet-50 architecture. Based on ImageNet pre-training, it can be fine-tuned using a polished surface dataset. Input: local image of the polished surface with a resolution of 224×224, normalized to the [0,1] interval; Output: roughness level or continuous roughness value.

[0143] Improved defect detection networks, such as YOLOv8 or Faster R-CNN-based object detection frameworks, can be used to identify and locate common grinding defects, such as over-grinding, under-grinding, vibration marks, burns, and scratches, and determine their severity levels.

[0144] S73. Based on the quality scores of each region, calculate the quality score of the workpiece to be polished and determine the quality grade.

[0145] Specifically, after conducting quality assessments for each region separately, the overall quality score can be calculated using the following weighted formula: in, The weight can be set based on area and process importance; The quality score (0~100) is given for the i-th region. A quality score above 90 is "Excellent", 80~90 is "Good", 70~80 is "Pass", and below 70 is "Fail".

[0146] S74. Based on the surface roughness assessment value and grinding defect status of each region, as well as the pre-set grinding target, determine the completion status of the grinding target task.

[0147] Specifically, the surface roughness assessment values ​​and grinding defect status of each area are compared with the pre-set grinding targets to determine the success or failure of the task. The actual detected roughness Ra and the number of defects can be compared with the preset target values ​​in the strategy: if Ra ≤ target value and there are no serious defects, the task is considered successful; if Ra > target value or there are serious defects, the task is considered a failure. If the grinding target task is not completed, proceed to step S75.

[0148] S75. Combine execution records and abnormal situations for cross-analysis to determine the cause of deviation, and take the quality score, quality grade, surface roughness assessment value, grinding defect status and grinding target task completion status of the workpiece to be ground as the evaluation result.

[0149] Specifically, when the quality does not meet the standards, a cross-analysis is conducted by combining the execution records and abnormal situations to preliminarily determine the cause of the deviation. For example, excessive force control deviation may be caused by force sensor calibration error or mismatch of impedance parameters; local defect concentration may be caused by tool wear or improper handling of abrupt changes in curved surfaces; and high overall roughness may be caused by tool grain size selection being too large or insufficient number of passes.

[0150] The quality assessment report provided in this embodiment is shown below: json { Assessment Report ID: QR-20260710-001 Generation Time: 2026-07-10T15:22:30Z Overall Assessment: { Overall quality score: 92.5 Quality Grade: Excellent "Goal achieved": true Roughness_Ra: {"Measured": 1.45, "Target": "≤1.6", "Judgment": "Meets Standard"}, Total Defects: 1 }, "Zoning Assessment": [ { "Region ID": 1, "Area Name": "Hypercurvature Head Area", Quality rating: 88.0 Roughness_Ra: 1.55, "Defect List": [{"Type": "Insufficient Wear", "Location": [125.6, 340.2, 85.0], "Severity": "Minor"}], "Assessment confidence level": 0.93 }, { "Region ID": 2, "Area Name": "Middle section of single-curvature fuselage", Quality rating: 95.0 Roughness_Ra: 1.38, "Defect List": [], "Assessment confidence level": 0.95 } ], Deviation Analysis: { Overall Deviation Trend: "Normal" "Main Deviation Items": "Underwear Defect in Area 1", Suggested improvements: "Add one more pass to area 1 or reduce the feed rate." } } S76. Collect the basic information of the workpiece, the semantic vector of the surface, the structured grinding strategy document, the execution record and the evaluation result as an experience set and store them in the historical experience case library.

[0151] Specifically, the complete data chain from perception to evaluation for this polishing task is structurally assembled into experience records and stored in an experience repository. Each polishing task generates one experience record, which can contain the following complete fields: json { "Experience ID": "EXP-20260710-001", "Timestamp": "2026-07-10T15:20:45Z", "Workpiece Information": { "Workpiece ID": "W20260710-001", Material: CFRP "Workpiece Type": "UAV Skin" }, "Strategy Information": { "Strategy ID": "POLICY-20260710-001", "Large Model Version": "GrindLLM-v2.3.1", "Generated parameters": { "Fd": [6.0, 10.0, 5.0], "Vd": [200, 300, 150]} }, "Sensory Data Summary": { "Surface Semantic Vector": { / * Step S1 outputs a summary * / }, Number of regions: 3 Total sanding area: 2.35 }, "Execution Data Summary": { "Actual force control accuracy": { "Maximum deviation": 1.8, "Average deviation": 0.6}, Actual processing time: 435. "Number of exceptions triggered": 2 }, "Evaluation Results": { Quality rating: 92.5 Roughness_Ra: 1.45, "Defect Statistics": { "Over-wearing": 0, "Under-wearing": 1, "Voice marks": 0, "Burns": 0}, "Goal achieved": true Quality Grade: "Excellent" }, Experience Tags: ["CFRP", "Hypercurvature", "Semi-finishing", "Success"] Reusability score: 0.85 } The experience base can be stored using a vector database (such as Milvus or Pinecone). Each experience record is indexed by a semantic vector generated by the model, which facilitates subsequent retrieval of similar cases based on similarity.

[0152] The complete data from this task is stored in an experience base, and a model fine-tuning mechanism is triggered periodically to continuously improve the model's process decision-making capabilities over time. The essence of this step is to realize the system's self-learning and self-evolution capabilities. Through an incremental update mechanism, the actual effects of each refinement are internalized as "experience" and transformed into searchable knowledge without changing the model's basic architecture.

[0153] In some embodiments of this application, when the experience base accumulates to a preset threshold, a model fine-tuning process is triggered, and the accumulated experience data is used to incrementally train the multimodal large model. Based on this, the method may further include: S81. Monitor the number of experience sets in the historical experience case library to determine whether the model fine-tuning conditions have been met.

[0154] Specifically, the system monitors the amount of data in the experience database and triggers fine-tuning when any of the following conditions are met: the number of newly added experience records is ≥100; the cumulative number of experience records is ≥500 (first trigger); or the time since the last fine-tuning is ≥30 days. If the model fine-tuning conditions are met, S82 is executed.

[0155] S82. Select a preset number of experience sets from the historical experience case library.

[0156] Specifically, high-quality experience records are selected from the experience base as fine-tuning datasets. The selection criteria can be set as: successful cases with a quality score ≥ 80 and a reusability score ≥ 0.7, including both successful and failed cases (including bias attribution analysis); the strategy parameters in the successful cases are slightly perturbed to generate simulation samples and enhance the model's generalization ability.

[0157] S83. Fine-tune the polishing strategy generation model using the selected experience set.

[0158] Specifically, fine-tuning can employ an incremental learning strategy, which incorporates new experiences while retaining the capabilities of the large model.

[0159] Edge devices have limited computing power and cannot support full model fine-tuning calculations. A collaborative architecture of "lightweight deployment at the edge + incremental training in the cloud" can be adopted. That is, a lightweight small model that has undergone knowledge distillation is deployed on the edge device to handle online inference; the full large model is deployed in the cloud and incrementally trained using new data in the experience base. After training, the updated capabilities are migrated to the edge through model compression and knowledge distillation, which not only ensures the real-time response capability of the edge, but also enables the continuous evolution of the model.

[0160] S84. Using a pre-created validation set, verify the improvement rate of process parameter recommendation accuracy, the improvement rate of anomaly cause judgment accuracy, and the improvement rate of strategy generation quality score of the fine-tuned grinding strategy generation model.

[0161] S85. Based on the improvement rate of process parameter recommendation accuracy, the improvement rate of anomaly cause judgment accuracy, and the improvement rate of strategy generation quality score, determine whether the fine-tuned grinding strategy generation model passes the evaluation.

[0162] Specifically, if the fine-tuned polishing strategy generation model passes the evaluation, then execute S86; if the fine-tuned polishing strategy generation model fails the evaluation, then execute S87.

[0163] S86. Deploy the fine-tuned polishing strategy generation model online.

[0164] S87. Revert to the previous version of the polishing strategy generation model and record the failure log.

[0165] The curved surface grinding apparatus provided in the embodiments of this application is described below. The curved surface grinding apparatus described below can be referred to in correspondence with the curved surface grinding method described above.

[0166] Figure 2 This is a schematic diagram of a curved surface grinding device provided in an embodiment of this application, with reference to... Figure 2 As shown, the device may include: The data acquisition module 100 is used to acquire the three-dimensional point cloud data, visual image data and user language commands of the workpiece to be polished; The digital surface model construction module 101 is used to construct a triangular mesh digital surface model of the workpiece to be polished based on three-dimensional point cloud data. Vertex parameter calculation module 102 is used to calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model; The region division module 103 is used to divide the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and to determine the region type label of each region. The surface semantic vector generation module 104 is used to encode the normal vector, curvature data and region type label of each vertex using a pre-trained lightweight graph neural network, and map them to a high-dimensional semantic embedding space to obtain a structured surface semantic vector. The polishing strategy generation module 105 is used to input a pre-trained polishing strategy generation model based on visual image data, user language instructions and structured surface semantic vectors to obtain a structured polishing strategy file. The structured polishing strategy file contains the differential process parameters corresponding to each region. The grinding execution module 106 is used to control the robot to grind the workpiece to be ground based on the structured grinding strategy file and generate an execution record. The multimodal data acquisition module 107 is used to acquire six-dimensional force signal data and tactile vibration signals in real time during the polishing process; The robot end-effector posture position correction module 108 is used to correct the robot end-effector posture position in real time based on six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex and the structured grinding strategy file. The abnormal situation monitoring module 109 is used to determine abnormal situations based on six-dimensional force signal data, tactile vibration signals, differential process parameters corresponding to each region, and pre-set abnormal thresholds, and to determine whether it is necessary to call the pre-trained adjustment decision model to adjust the grinding strategy based on the abnormal situation. The historical adjustment decision retrieval module 110 is used to search for similar adjustment decision records from a pre-created historical experience case library when it is necessary to call a pre-trained adjustment decision model to adjust the polishing strategy, based on the abnormal situation and the structured surface semantic vector corresponding to the current area. The prompt word determination module 111 is used to determine prompt words based on six-dimensional force signal data, tactile vibration signals, the structured surface semantic vector corresponding to the current area, and similar adjustment decision records; The adjustment instruction generation module 112 is used to input prompt words into the adjustment decision model to obtain adjustment instructions. The adjustment decision model is configured to have the ability to obtain the current state context matrix based on six-dimensional force signal data and tactile vibration signals, obtain the anomaly cause reasoning result based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain adjustment instructions based on the anomaly cause reasoning result and similar adjustment decision records. The grinding strategy optimization and adjustment module 113 is used to adjust the structured grinding strategy file according to the adjustment instructions, obtain the adjusted structured grinding strategy instructions, and return to execute the steps of controlling the robot to grind the workpiece to be ground based on the structured grinding strategy file until the grinding is completed.

[0167] This application provides a surface polishing method, comprising: a data acquisition module 100 for acquiring three-dimensional point cloud data, visual image data, and user language commands of the workpiece to be polished; a digital surface model construction module 101 for constructing a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data; a vertex parameter calculation module 102 for calculating the normal vector and curvature data of each vertex in the triangular mesh digital surface model; a region division module 103 for dividing the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and determining the region type label of each region; and a surface semantic vector generation module 104 for using a pre-trained lightweight graph neural network. The network encodes the normal vectors, curvature data, and region type labels of each vertex, mapping them to a high-dimensional semantic embedding space to obtain structured surface semantic vectors. A polishing strategy generation module 105, based on visual image data, user language commands, and structured surface semantic vectors, inputs a pre-trained polishing strategy generation model to obtain a structured polishing strategy file. This file contains differential process parameters corresponding to each region. A polishing execution module 106, based on the structured polishing strategy file, controls the robot to polish the workpiece and generates an execution record. A multimodal data acquisition module 107 is used to collect six-dimensional force signal data and tactile vibration signals during the polishing process in real time. The robot end effector... The end-effector position correction module 108 is used to correct the robot's end-effector position in real time based on six-dimensional force signal data, differential process parameters corresponding to each region, normal vectors of each vertex, and structured grinding strategy files. The anomaly monitoring module 109 is used to identify anomalies based on six-dimensional force signal data, tactile vibration signals, differential process parameters corresponding to each region, and pre-set anomaly thresholds, and based on the anomalies, determine whether it is necessary to call a pre-trained adjustment decision model to adjust the grinding strategy. The historical adjustment decision retrieval module 110 is used to retrieve historical adjustment decisions based on the anomalies and the structured surface semantic vectors corresponding to the current region when it is necessary to call a pre-trained adjustment decision model to adjust the grinding strategy. The system searches for similar adjustment decision records from a pre-created historical experience case library; the prompt word determination module 111 is used to determine prompt words based on six-dimensional force signal data, tactile vibration signals, the structured surface semantic vector corresponding to the current region, and similar adjustment decision records; the adjustment instruction generation module 112 is used to input the prompt words into the adjustment decision model to obtain adjustment instructions. The adjustment decision model is configured to have the ability to obtain the current state context matrix based on six-dimensional force signal data and tactile vibration signals, obtain the anomaly cause reasoning result based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain adjustment instructions based on the anomaly cause reasoning result and similar adjustment decision records.The grinding strategy optimization and adjustment module 113 is used to adjust the structured grinding strategy file according to the adjustment instructions, obtain the adjusted structured grinding strategy instructions, and return to execute the steps of controlling the robot to grind the workpiece based on the structured grinding strategy file until grinding is completed. This application generates a grinding strategy based on surface semantic features and adjusts the grinding strategy in real time based on the state data during the grinding process, controlling the robot to achieve surface grinding.

[0168] Optionally, the digital surface model building module 101 may perform the process of building a triangular mesh digital surface model of the workpiece to be polished based on three-dimensional point cloud data, which may include: For each point, calculate the average distance from all points in its neighborhood to that point; Points whose average distance exceeds the preset global average distance threshold are identified as outliers and removed. The 3D point cloud space is divided into voxel grids of a specified size. The 3D point cloud data, after removing outlier noise, is simplified by retaining only one representative point in each grid. Based on the simplified 3D point cloud data, a continuous triangular mesh digital surface model of the workpiece to be polished is constructed.

[0169] Optionally, the curvature data includes average curvature. The region partitioning module 103 performs the process of partitioning the triangular mesh digital surface model into regions based on the normal vectors and curvature data of each vertex, and determining the region type label for each region. This process may include: From the vertices whose region type labels have not been determined, select the vertex corresponding to the minimum average curvature as the current seed point, and determine the region type label of the region where the current seed point is located based on the normal vector and curvature data of the current seed point. The neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset threshold for the angle between their normal vectors and a curvature difference that is less than a preset threshold for the curvature difference are included in the region where the current seed point is located. Determine if there are any newly added neighboring vertices in the region where the current seed point is located; If there are newly added neighboring vertices in the region where the current seed point is located, then the newly added neighboring vertices in the region where the current seed point is located will be used as the new current seed point in turn, and the process will return to execute the step of adding the neighboring vertices with undetermined region type labels that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold to the region where the current seed point is located. If no new neighboring vertices are added to the region where the current seed point is located, return to the step of selecting the vertex with the minimum average curvature from the vertices whose region type labels have not been determined as the current seed point.

[0170] Optionally, the polishing strategy generation model includes: an input layer, a multimodal semantic fusion layer, a polishing state understanding and process reasoning layer, a partitioning strategy and process parameter generation layer, a baseline polishing path generation layer, and a structured polishing strategy file output layer. The device also includes a polishing strategy generation model training module for training the model. The process of training the polishing strategy generation model may include: The input layer acquires visual image data, user language commands, and structured surface semantic vectors. Through the multimodal semantic fusion layer, the input visual image data, user language commands, and structured surface semantic vectors are mapped to a unified semantic embedding space, thereby achieving adaptive fusion of cross-modal features and obtaining fused semantic features. By understanding the grinding status and reasoning the process, based on the fused semantic features, the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each area, and the current grinding stage are determined. By using a partitioning strategy and a process parameter generation layer, the processing order between regions is generated based on the comprehensive processing difficulty level of each region and the spatial adjacency between regions. Based on the processing order between regions and the fused semantic features, the differentiated process parameters corresponding to each region are generated. Through a baseline grinding path generation layer, for each region, based on the region type label, different path generation strategies are adopted to create target grinding points for each region. According to a pre-defined objective function, the smoothness of the target grinding points is optimized, and the grinding path for each region is determined. Based on the machining sequence between regions and the corresponding grinding paths, the transition relationship between regions is determined. The target grinding point includes spatial coordinates, tool posture, desired contact force, and desired speed. The structured grinding strategy file output layer generates a structured grinding strategy file based on the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each region, the current grinding stage, the differentiated process parameters corresponding to each region, the path generation strategy, the target grinding points corresponding to each region, and the transition relationship between regions.

[0171] Optionally, the differential process parameters include the target normal contact force. The robot end-effector posture position correction module 108 executes a process based on six-dimensional force signal data, differential process parameters corresponding to each region, normal vectors of each vertex, and structured grinding strategy files to correct the robot end-effector posture position in real time. This process may include: Determine the normal vector of the current contact point based on the normal vectors of the three vertices on the triangular surface where the current contact point is located; Determine the current normal contact force based on the six-dimensional force signal data and normal vector at the current contact point; Calculate the deviation between the current normal contact force and the target normal contact force in real time; Using the deviation value and the differential process parameters corresponding to the current region, calculate the robot end position correction amount, and based on the robot end position correction amount, correct the robot end position in the normal direction of the current contact point; The path tangent direction of the current contact point is determined based on the structured polishing strategy file; Based on the normal vector of the current contact point and the tangent direction of the path, determine the desired posture of the robot's end effector; Acquire the current posture data of the robot's end effector and calculate the current posture error; Based on the current attitude error and angular velocity error, a robot end-effector attitude correction command is generated, and the robot end-effector attitude is corrected in the tangent plane direction of the surface where the current contact point is located based on the robot end-effector attitude correction command.

[0172] Optionally, the abnormal situation monitoring module 109 executes a process to determine abnormal situations based on six-dimensional force signal data, tactile vibration signals, differential process parameters corresponding to each region, and pre-set abnormal thresholds, which may include: Based on six-dimensional force signal data, a real-time force perception semantic vector is generated. Vibration feature vectors are extracted based on tactile vibration signals; By using force perception semantic vectors with preset periods and corresponding vibration feature vectors, and concatenating them according to time series, a state context matrix is ​​constructed. Based on the state context matrix, the different process parameters corresponding to each region, and the pre-set anomaly threshold, multi-dimensional anomaly detection is performed to determine the abnormal situation.

[0173] Optionally, the state context matrix includes the actual normal contact force, actual vibration amplitude, actual root mean square value of acoustic emission, actual force signal variance, and actual grinding area. The differential process parameters include the target normal contact force. The abnormal situation monitoring module 109 performs multi-dimensional abnormality detection based on the state context matrix, the differential process parameters corresponding to each region, and pre-set abnormality thresholds to determine the abnormal situation. This process may include: If the difference between the actual normal contact force and the target normal contact force is greater than the preset abnormal threshold for normal contact force, and the duration is greater than the preset deviation time threshold, then it is determined to be a force deviation warning. If the rate of change of the difference between the actual normal contact force and the target normal contact force is greater than the preset threshold for the rate of change of the deviation of the normal contact force, it is determined as a force change warning. If the actual vibration amplitude is greater than the preset vibration amplitude threshold, it is determined as a tremor risk warning. If the actual root mean square value of acoustic emission is greater than the preset root mean square value threshold of acoustic emission, it is determined as a tool passivation or burn warning. If the actual force signal variance is greater than the preset force signal variance, it is determined to be a contact instability warning. If the actual polishing area differs from the target polishing area, it will be considered a parameter update warning. Record the warning situation and the number of warnings as abnormal situations.

[0174] Optionally, the device may also include: The multimodal data real-time acquisition module is used to acquire six-dimensional force signal data and tactile vibration signals in real time during the polishing process; The abnormal situation change monitoring module is used to monitor the changing trend of abnormal situations based on six-dimensional force signal data and tactile vibration signals; The adjustment judgment module is used to determine the adjustment is effective and record it as a successful experience when the abnormal indicator corresponding to the abnormal situation gradually converges to the target value. When the abnormal indicator corresponding to the abnormal situation fails to converge to the target value or a new abnormal situation occurs, the adjustment is determined to be invalid, the current adjustment instruction is generated as an adjustment failure result, and the current adjustment instruction as an adjustment failure result is added to the prompt word to obtain a new prompt word. The module then returns to the execution step of inputting the prompt word into the adjustment decision model to obtain the adjustment instruction. If the adjustment is determined to be invalid twice consecutively, the safety protection mechanism is triggered and the failure is recorded as a failure experience.

[0175] Optionally, the device may also include: The post-grinding image acquisition module is used to acquire post-grinding image data of the workpiece to be ground; The evaluation module is used to determine the surface roughness evaluation value, grinding defects, and quality score of each area based on the image data after grinding. The quality scoring module is used to calculate the quality score of the workpiece to be polished based on the quality score of each region, and to determine the quality level. The grinding target completion judgment module is used to judge the completion status of grinding target tasks based on the surface roughness evaluation value and grinding defect status of each area, as well as the pre-set grinding target. If it is not completed, it performs cross-analysis by combining execution records and abnormal situations to determine the cause of deviation, and uses the quality score, quality grade, surface roughness evaluation value, grinding defect status and grinding target task completion status of the workpiece to be ground as the evaluation result. The experience set storage module is used to store the basic information of the workpiece, the semantic vector of the surface, the structured grinding strategy file, the execution record and the evaluation result as an experience set in the historical experience case library.

[0176] Optionally, the device may also include: The model fine-tuning module monitors the number of experience sets in the historical experience case library to determine if the model fine-tuning conditions are met. If met, it selects a preset number of experience sets from the historical experience case library. The selected experience sets are then used to fine-tune the polishing strategy generation model. A pre-created validation set is used to verify the improvement rates of process parameter recommendation accuracy, anomaly cause judgment accuracy, and strategy generation quality score of the fine-tuned polishing strategy generation model. Based on these improvements, the model is evaluated to determine if it passes the evaluation. If it passes, the fine-tuned model is deployed online. If it fails, it reverts to the previous version of the polishing strategy generation model and records the failure log.

[0177] This application also provides a surface grinding device. Figure 3 The hardware structure block diagram of the curved surface grinding equipment is shown. (Refer to...) Figure 3 The hardware structure of the curved surface grinding equipment may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4; In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4; Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device; The memory stores a program, which the processor can call. The program is used to implement the various processing steps in the aforementioned surface polishing method.

[0178] This application also provides a computer-readable storage medium that stores a program suitable for processor execution, the program being used to implement various processing steps in the aforementioned surface polishing method.

[0179] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0180] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined with each other, and the same or similar parts can be referred to each other.

[0181] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for polishing curved surfaces, characterized in that, include: Acquire 3D point cloud data, visual image data, and user language commands for the workpiece to be polished; Based on the three-dimensional point cloud data, a triangular mesh digital surface model of the workpiece to be polished is constructed. Calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model; Based on the normal vectors and curvature data of each vertex, the triangular mesh digital surface model is divided into regions, and the region type label of each region is determined. By using a pre-trained lightweight graph neural network, the normal vectors, curvature data and region type labels of each vertex are encoded and mapped to a high-dimensional semantic embedding space to obtain structured surface semantic vectors. Based on the visual image data, the user language instructions, and the structured surface semantic vector, a pre-trained polishing strategy generation model is input to obtain a structured polishing strategy file, which contains differential process parameters corresponding to each region. Based on the structured grinding strategy file, the robot is controlled to grind the workpiece to be ground and an execution record is generated. Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process; Based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex, and the structured grinding strategy file, the robot end effector position is corrected in real time. Based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, abnormal situations are identified, and based on the abnormal situations, it is determined whether it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy. If necessary, based on the aforementioned abnormal situation and the structured surface semantic vector corresponding to the current region, similar adjustment decision records are searched from a pre-created historical experience case library; Based on the six-dimensional force signal data, the tactile vibration signal, the structured surface semantic vector corresponding to the current region, and the similar adjustment decision records, the prompt words are determined; The prompt word is input into the adjustment decision model to obtain the adjustment instruction. The adjustment decision model is configured to have the ability to obtain the current state context matrix based on the six-dimensional force signal data and the tactile vibration signal, obtain the anomaly cause reasoning result based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain the adjustment instruction based on the anomaly cause reasoning result and the similar adjustment decision record. Based on the adjustment instructions, the structured grinding strategy file is adjusted accordingly to obtain the adjusted structured grinding strategy instructions. Then, the process of controlling the robot to grind the workpiece based on the structured grinding strategy file is returned to be executed until the grinding is completed.

2. The method according to claim 1, characterized in that, The process of constructing a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data includes: For each point, calculate the average distance from all points in its neighborhood to that point; Points whose average distance exceeds the preset global average distance threshold are identified as outliers and removed. The 3D point cloud space is divided into voxel grids of a specified size. The 3D point cloud data, after removing outlier noise, is simplified by retaining only one representative point in each grid. Based on the simplified 3D point cloud data, a continuous triangular mesh digital surface model of the workpiece to be polished is constructed.

3. The method according to claim 1, characterized in that, The curvature data includes average curvature. Based on the normal vectors and curvature data of each vertex, the triangular mesh digital surface model is divided into regions, and the region type label of each region is determined, including: From the vertices whose region type labels have not been determined, select the vertex corresponding to the minimum average curvature as the current seed point, and determine the region type label of the region where the current seed point is located based on the normal vector and curvature data of the current seed point. The neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold are included in the region where the current seed point is located. Determine whether there are any newly added neighboring vertices in the region where the current seed point is located; If there are newly added neighborhood vertices in the region where the current seed point is located, then the neighboring vertices in the region where the current seed point is located will be used as new current seed points in turn, and the process will return to the step of adding the neighboring vertices of the undetermined region type label that have an angle between their normal vector and the current seed point that is less than a preset normal vector angle threshold and a curvature difference that is less than a preset curvature difference threshold to the region where the current seed point is located. If no new vertex is added to the region where the current seed point is located, then return to the step of selecting the vertex with the minimum average curvature from the vertices whose region type labels have not been determined as the current seed point.

4. The method according to claim 1, characterized in that, The polishing strategy generation model includes: an input layer, a multimodal semantic fusion layer, a polishing state understanding and process reasoning layer, a partitioning strategy and process parameter generation layer, a baseline polishing path generation layer, and a structured polishing strategy file output layer. The visual image data, the user language instructions, and the structured surface semantic vector are obtained through the input layer. The multimodal semantic fusion layer maps the input visual image data, user language commands, and structured surface semantic vectors to a unified semantic embedding space, thereby achieving adaptive fusion of cross-modal features and obtaining fused semantic features. Through the grinding state understanding and process reasoning layer, based on the fused semantic features, the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each area, and the current grinding stage are determined. Through the partitioning strategy and process parameter generation layer, the processing order between regions is generated based on the comprehensive processing difficulty level of each region and the spatial adjacency relationship between each region. Based on the processing order between regions and the fused semantic features, the differentiated process parameters corresponding to each region are generated. Through the aforementioned reference grinding path generation layer, for each region, based on the region type label corresponding to each region, different path generation strategies are adopted to create target grinding points corresponding to each region. According to the pre-set objective function, the smoothness of the target grinding points is optimized, and the grinding path corresponding to each region is determined. Based on the processing sequence between the regions and the grinding path corresponding to each region, the transition relationship between regions is determined. The target grinding point includes spatial coordinates, tool posture, desired contact force, and desired speed. The structured grinding strategy file output layer generates a structured grinding strategy file based on the material and physical properties of the workpiece to be ground, the comprehensive processing difficulty level of each region, the current grinding stage, the differentiated process parameters corresponding to each region, the path generation strategy, the target grinding point number corresponding to each region, and the transition relationship between regions.

5. The method according to claim 1, characterized in that, The differential process parameters include the target normal contact force. The real-time correction of the robot's end effector posture position based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vectors of each vertex, and the structured grinding strategy file includes: The normal vector of the current contact point is determined based on the normal vectors of the three vertices on the triangular surface where the current contact point is located. Determine the current normal contact force based on the six-dimensional force signal data and normal vector at the current contact point; The deviation between the current normal contact force and the target normal contact force is calculated in real time. Using the deviation value and the differential process parameters corresponding to the current region, calculate the robot end position correction amount, and based on the robot end position correction amount, correct the robot end position in the normal direction of the current contact point; The path tangent direction of the current contact point is determined based on the structured polishing strategy file; Based on the normal vector of the current contact point and the tangent direction of the path, the desired posture of the robot end effector is determined; Acquire the current posture data of the robot's end effector and calculate the current posture error; Based on the current attitude error and angular velocity error, a robot end-effector attitude correction command is generated, and the robot end-effector attitude is corrected in the tangent plane direction of the surface where the current contact point is located based on the robot end-effector attitude correction command.

6. The method according to claim 1, characterized in that, The process of determining abnormal situations based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and a pre-set abnormal threshold includes: Based on the six-dimensional force signal data, a real-time force perception semantic vector is generated; Based on the tactile vibration signal, a vibration feature vector is extracted; By using force perception semantic vectors with preset periods and corresponding vibration feature vectors, and concatenating them according to time series, a state context matrix is ​​constructed. Based on the state context matrix, the differential process parameters corresponding to each region, and the pre-set anomaly threshold, multi-dimensional anomaly detection is performed to determine the abnormal situation.

7. The method according to claim 6, characterized in that, The state context matrix includes the actual normal contact force, actual vibration amplitude, actual root mean square value of acoustic emission, actual force signal variance, and actual grinding area. The differential process parameters include the target normal contact force. Based on the state context matrix, the differential process parameters corresponding to each region, and a pre-set anomaly threshold, multi-dimensional anomaly detection is performed to determine anomalies, including: If the difference between the actual normal contact force and the target normal contact force is greater than the preset abnormal threshold for normal contact force, and the duration is greater than the preset deviation time threshold, then it is determined to be a force deviation warning. If the rate of change of the difference between the actual normal contact force and the target normal contact force is greater than the preset threshold for the rate of change of the normal contact force deviation, it is determined as a force change warning. If the actual vibration amplitude is greater than the preset vibration amplitude threshold, it is determined to be a tremor risk warning; If the actual root mean square value of acoustic emission is greater than the preset root mean square value threshold of acoustic emission, it is determined as a tool passivation or burn warning. If the actual force signal variance is greater than the preset force signal variance, it is determined to be a contact instability warning; If the actual polishing area differs from the target polishing area, it will be considered a parameter update warning. Record the warning situation and the number of warnings as abnormal situations.

8. The method according to any one of claims 1-7, characterized in that, After receiving the adjusted structured grinding strategy instructions and controlling the robot to grind the workpiece based on the adjusted structured grinding strategy file, the process also includes: Real-time acquisition of six-dimensional force signal data and tactile vibration signals during the polishing process; Based on the six-dimensional force signal data and the tactile vibration signal, monitor the changing trend of the abnormal situation; If the abnormal indicators corresponding to the abnormal situation gradually converge to the target value, the adjustment is deemed effective and recorded as a successful experience. If the abnormal indicator corresponding to the abnormal situation does not gradually converge to the target value or a new abnormal situation occurs, the adjustment is determined to be invalid, the current adjustment instruction is generated as an adjustment failure result, and the current adjustment instruction as an adjustment failure result is added to the prompt word to obtain a new prompt word. The process is then returned to the step of inputting the prompt word into the adjustment decision model to obtain the adjustment instruction. If the adjustment is determined to be invalid twice in a row, the safety protection mechanism is triggered and recorded as a failure experience.

9. The method according to any one of claims 1-7, characterized in that, After polishing, the process also includes: Collect image data of the workpiece to be polished after polishing; Based on the polished image data, the surface roughness assessment value, polishing defects, and quality score of each region are determined. Based on the quality scores of each region, the quality score of the workpiece to be polished is calculated, and the quality grade is determined. Based on the surface roughness assessment values ​​and grinding defects of each region, as well as the pre-set grinding targets, the completion status of the grinding target tasks is judged. If not completed, the execution record and the abnormal situation are cross-analyzed to determine the cause of the deviation, and the quality score, quality grade, surface roughness evaluation value, grinding defect situation and grinding target task completion status of the workpiece to be ground are used as the evaluation result. The basic information of the workpiece, the semantic vector of the surface, the structured grinding strategy file, the execution record, and the evaluation result are stored as an experience set in the historical experience case library.

10. The method according to claim 9, characterized in that, Also includes: Monitor the number of experience sets in the historical experience case library to determine whether the model fine-tuning conditions have been met. If the target is reached, a preset number of experience sets will be selected from the historical experience case library. The polishing strategy generation model was fine-tuned using the selected experience set; Using a pre-created validation set, we validated the improvement rates of process parameter recommendation accuracy, anomaly cause judgment accuracy, and strategy generation quality score of the fine-tuned grinding strategy generation model. Based on the improvement rate of the recommended process parameters, the improvement rate of the accuracy of the judgment of the cause of the anomaly, and the improvement rate of the quality score of the strategy generation, it is determined whether the fine-tuned polishing strategy generation model passes the evaluation. If approved, the refined polishing strategy generation model will be deployed online. If it fails, it will revert to the previous version of the polishing strategy generation model and record the failure log.

11. A curved surface grinding device, characterized in that, include: The data acquisition module is used to acquire the 3D point cloud data, visual image data and user language commands of the workpiece to be polished; A digital surface model construction module is used to construct a triangular mesh digital surface model of the workpiece to be polished based on the three-dimensional point cloud data. The vertex parameter calculation module is used to calculate the normal vector and curvature data of each vertex in the triangular mesh digital surface model; The region division module is used to divide the triangular mesh digital surface model into regions based on the normal vector and curvature data of each vertex, and determine the region type label of each region. The surface semantic vector generation module is used to encode the normal vector, curvature data and region type label of each vertex using a pre-trained lightweight graph neural network, and map them to a high-dimensional semantic embedding space to obtain a structured surface semantic vector. The polishing strategy generation module is used to input a pre-trained polishing strategy generation model based on the visual image data, the user language instructions and the structured surface semantic vector to obtain a structured polishing strategy file, wherein the structured polishing strategy file contains differential process parameters corresponding to each region. The grinding execution module is used to control the robot to grind the workpiece to be ground based on the structured grinding strategy file and generate an execution record. The multimodal data acquisition module is used to acquire six-dimensional force signal data and tactile vibration signals in real time during the polishing process; The robot end effector posture position correction module is used to correct the robot end effector posture position in real time based on the six-dimensional force signal data, the differential process parameters corresponding to each region, the normal vector of each vertex and the structured grinding strategy file. An abnormal situation monitoring module is used to determine abnormal situations based on the six-dimensional force signal data, the tactile vibration signal, the differential process parameters corresponding to each region, and the pre-set abnormal threshold, and to determine whether it is necessary to call the pre-trained adjustment decision model to adjust the polishing strategy based on the abnormal situation. The historical adjustment decision retrieval module is used to search for similar adjustment decision records from a pre-created historical experience case library when it is necessary to call a pre-trained adjustment decision model to adjust the polishing strategy, based on the abnormal situation and the structured surface semantic vector corresponding to the current region. The prompt word determination module is used to determine prompt words based on the six-dimensional force signal data, the tactile vibration signal, the structured surface semantic vector corresponding to the current region, and the similar adjustment decision records; An adjustment instruction generation module is used to input prompt words into the adjustment decision model to obtain adjustment instructions. The adjustment decision model is configured to have the ability to obtain a current state context matrix based on the six-dimensional force signal data and the tactile vibration signal, obtain anomaly cause inference results based on the current state context matrix and the structured surface semantic vector corresponding to the current region, and obtain adjustment instructions based on the anomaly cause inference results and the similar adjustment decision records. The grinding strategy optimization and adjustment module is used to adjust the structured grinding strategy file according to the adjustment instructions to obtain the adjusted structured grinding strategy instructions, and return to execute the steps of controlling the robot to grind the workpiece to be ground based on the structured grinding strategy file until the grinding is completed.

12. A curved surface grinding device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement the various steps of the surface polishing method as described in any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the various steps of the surface polishing method as described in any one of claims 1-10.

Images